A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 15, 09:45 UTC
[Minor] Rename quantization nvfp4 to modelopt_fp4 (#18356)
Signed-off-by: mgoin <mgoin64@gmail.com>
Michael Goin committed
May 20, 2025 at 12:08 UTC
f4a8a3746575ea62fd3e0b4bd29ebd782582836d
4 files changed
+7
-7
tests/models/quantization/test_nvfp4.py
+3
-3
@@ -41,8 +41,8 @@ EXPECTED_STRS_MAP = {
41
reason=
42
"Prevent unstable test based on golden strings from breaking the build "
43
" and test input model being too large and hanging the system.")
44
-@pytest.mark.skipif(not is_quant_method_supported("nvfp4"),
45
- reason="nvfp4 is not supported on this GPU type.")
44
+@pytest.mark.skipif(not is_quant_method_supported("modelopt_fp4"),
45
+ reason="modelopt_fp4 is not supported on this GPU type.")
46
@pytest.mark.parametrize("model_name", MODELS)
47
def test_models(example_prompts, model_name) -> None:
48
model = LLM(
@@ -50,7 +50,7 @@ def test_models(example_prompts, model_name) -> None:
50
max_model_len=MAX_MODEL_LEN,
51
trust_remote_code=True,
52
enforce_eager=True,
53
- quantization="nvfp4",
53
+ quantization="modelopt_fp4",
54
)
55
56
tokenizer = AutoTokenizer.from_pretrained(model_name)
vllm/config.py
+1
-1
@@ -824,7 +824,7 @@ class ModelConfig:
824
optimized_quantization_methods = [
825
"fp8", "marlin", "modelopt", "gptq_marlin_24", "gptq_marlin",
826
"awq_marlin", "fbgemm_fp8", "compressed-tensors", "experts_int8",
827
- "quark", "nvfp4", "bitblas", "gptq_bitblas"
827
+ "quark", "modelopt_fp4", "bitblas", "gptq_bitblas"
828
]
829
if self.quantization is not None:
830
self.quantization = cast(QuantizationMethods,
vllm/model_executor/layers/quantization/__init__.py
+2
-2
@@ -14,7 +14,7 @@ QuantizationMethods = Literal[
14
"ptpc_fp8",
15
"fbgemm_fp8",
16
"modelopt",
17
- "nvfp4",
17
+ "modelopt_fp4",
18
"marlin",
19
"bitblas",
20
"gguf",
@@ -120,7 +120,7 @@ def get_quantization_config(quantization: str) -> type[QuantizationConfig]:
120
"fp8": Fp8Config,
121
"fbgemm_fp8": FBGEMMFp8Config,
122
"modelopt": ModelOptFp8Config,
123
- "nvfp4": ModelOptNvFp4Config,
123
+ "modelopt_fp4": ModelOptNvFp4Config,
124
"marlin": MarlinConfig,
125
"bitblas": BitBLASConfig,
126
"gguf": GGUFConfig,
vllm/model_executor/layers/quantization/modelopt.py
+1
-1
@@ -192,7 +192,7 @@ class ModelOptNvFp4Config(QuantizationConfig):
192
193
@classmethod
194
def get_name(cls) -> QuantizationMethods:
195
- return "nvfp4"
195
+ return "modelopt_fp4"
196
197
@classmethod
198
def get_supported_act_dtypes(cls) -> list[torch.dtype]: